Upcoming SlideShare
×

# Kalman filter implimention in mathlab

5,713 views

Published on

Introduce the kalman filter

• Full Name
Comment goes here.

Are you sure you want to Yes No
• Hi . exactly , where we can find 'format' folder to see text files ?

Are you sure you want to  Yes  No
• Hi! Where can I find the 'format' folder from where I can see the format of the data in text files?

Are you sure you want to  Yes  No

### Kalman filter implimention in mathlab

1. 1. Universität Stuttgart Institute of Geodesy Kalman Filtering Implementation with Matlab Study Report in the Field of Study Geodesy and Geoinformatics at Universität Stuttgart Rachel Kleinbauer Helsinki, November 2004 Adviser:Prof. Dr.-Ing.habil. Dr.tech.h.c.mult Dr.-Ing.E.h.mult Erik W. Grafarend Universität Stuttgart Prof. Martin Vermeer Helsinki University of Technology
2. 2. AbstractIn 1960 and 1961 Rudolf Emil Kalman published his papers on a recursivepredictive filter that is based on the use of state space techniques and recursivealgorithms and therewith he revolutionized the field of estimation. Since that timethe so-called Kalman filter has been the subject of extensive research andapplication.The Kalman filter estimates the state of a dynamic system, even if the preciseform of the system is unknown. The filter is very powerful in the sense that itsupports estimations of past, present, and even future states.Within the scope of this study thesis it was the task to program a Kalman filter inMatlab. The intention is to give the students of the course “Methods ofNavigation” an understanding of the Kalman filter by providing them with itspractical aspects.The composition includes a description of the standard Kalman filter and itsalgorithm with the two main steps, the prediction step and the correction step.Furthermore the extended Kalman filter is discussed, which represents theconversion of the Kalman filter to nonlinear systems. In the end the program wasexecuted to calculate the orbit of a geostationary satellite as an example.Keywords: filter, Kalman gain matrix, prediction, dynamic model, state vectorSchlüsselwörter: Filter, Kalman Korrekturmatrix, Vorhersage, dynamisches Modell, Zustandsvekor
3. 3. Contents1 Introduction....................................................................................................3 1.1 Motivation....................................................................................................3 1.2 Preview ........................................................................................................42 General Information......................................................................................5 2.1 Kalman Filter ...............................................................................................5 2.2 State Vector..................................................................................................6 2.3 Dynamic Model ...........................................................................................7 2.4 Observation Model ......................................................................................73 Kalman Filter Algorithm ..............................................................................9 3.1 Prediction .....................................................................................................9 3.2 Correction ..................................................................................................114 Extended Kalman Filter..............................................................................13 4.1 Nonlinear Systems .....................................................................................13 4.2 Prediction ...................................................................................................14 4.3 Correction ..................................................................................................145 Summary.......................................................................................................166 Implementation ............................................................................................18 6.1 Extraction of the Data ................................................................................19 6.2 Prediction ...................................................................................................19 6.3 Correction ..................................................................................................23 1
4. 4. 7 Example ........................................................................................................26 7.1 Dynamic Model and Initial Values............................................................26 7.2 Observation Model and Observations........................................................29 7.3 Results........................................................................................................298 Conclusion and Prospect .............................................................................34References .............................................................................................................35 2
5. 5. IntroductionChapter 1Introduction1.1 MotivationNavigation systems become standard, as they are sold more and more, particularlywith new cars. Most of these navigation systems use no longer only the GlobalPositioning System (GPS) but also an inertial navigation system (INS) to help thedriver find his way. Together the two systems complement each other and permitimproved navigation accuracy and reliability especially when GPS is degraded orinterrupted for example because of buildings or tunnels.And for this application the Kalman filter provides the basis. It constitutes a toolfor correcting the predicted INS trajectory with GPS measurements.Also the determination of a reference orbit for these GPS satellites and correctingit with the data from the GPS control stations is a very important application ofthe Kalman filter.But these are only two examples of the wide variety of fields where Kalmanfiltering plays an important role. The application areas span from aerospace, tomarine navigation, demographic modelling, weather science, manufacturing andmany others. Because the Kalman filter is very effective and useful for such alarge class of problems, it has been subject of extensive research.Within the scope of this study thesis I programmed a Kalman filter in Matlab thatis meant to give the students an understanding of the Kalman filter by providingthem with its practical aspects. This programme will be used in the course„Methods of Navigation“ that the students can discover how the Kalman filterworks by observing it in action. 3
6. 6. Introduction1.2 PreviewAt the beginning, in chapter two the basic information is provided. The question“What is a Kalman filter” is answered and the basic components of the Kalmanfilter are explained.The third chapter is about the formulas of the standard Kalman filter, which is alinear filter. It describes the two main steps of the Kalman filter.The fourth chapter shows how these formulas are transferred to nonlinear systems,which leads to the so-called Extended Kalman filter.In chapter five the essential formulas of both the standard Kalman filter and theExtended Kalman filter are summarized in a table.Chapter six describes the implementation of the Kalman filter in Matlab withsome illustrative sections of the Matlab source code.The programmed Kalman filter is applied in chapter 7 to the example of ageostationary orbit.And finally chapter 8 represents the closing with conclusions and prospects. 4
7. 7. General InformationChapter 2General Information2.1 Kalman FilterRudolf Emil Kalman was born in Budapest, Hungary, on May 19, 1930. He hadthe idea of the Kalman filter for the first time in in the year 1958. In 1960 and1961 he published his papers on the Kalman filter and therewith he revolutionizedthe field of estimation.The Kalman filter is a recursive predictive filter that is based on the use of statespace techniques and recursive algorithms. It estimates the state of a dynamicsystem. This dynamic system can be disturbed by some noise, mostly assumed aswhite noise. To improve the estimated state the Kalman filter uses measurementsthat are related to the state but disturbed as well.Thus the Kalman filter consists of two steps: 1. the prediction 2. the correctionIn the first step the state is predicted with the dynamic model. In the second step itis corrected with the observation model, so that the error covariance of theestimator is minimized. In this sense it is an optimal estimator. 5
8. 8. General InformationThis procedure is repeated for each time step with the state of the previous timestep as initial value. Therefore the Kalman filter is called a recursive filter. Figure 1: circuit of the Kalman filterThe basic components of the Kalman filter are the state vector, the dynamic modeland the observation model, which are described below.2.2 State VectorThe state vector contains the variables of interest. It describes the state of thedynamic system and represents its degrees of freedom. The variables in the statevector cannot be measured directly but they can be inferred from values that aremeasurable.Elements of the state vector can be e.g. position, velocity, orientation angles, etc.A very simple example is a train that is driving with a constant velocity on astraight rail. In this case the train has two degrees of freedom, the distance s andthe velocity s = v . Thus the state vector is ⎡s⎤ x=⎢ ⎥ ⎣v ⎦ 6
9. 9. General InformationThe state vector has two values at the same time, that is the a priori value, thepredicted value before the update, and the a posteriori value, the corrected value −after the update. In the following the a priori value is marked by x and the a +posteriori value by x .2.3 Dynamic ModelThe dynamic model describes the transformation of the state vector over time. Itcan usually be represented by a system of differential equations. d x (t ) = x(t ) = f ( x (t ), m(t )) (2.1) dtIn the linear case this can easily be rewritten as x(t ) = F ⋅ x(t ) + n(t ) (2.2)where F is the dynamic matrix and is constant, x(t ) is the state vector and n(t )is the dynamic noise which is usually assumed as white noise and has thecovariance matrix Q (t ) .There are systems that cannot be modelled by differential equations, but thesesystems are not discussed in this thesis.2.4 Observation ModelThe observation model represents the relationship between the state and themeasurements. In the linear case the measurements can be described by a systemof linear equations, which depend on the state variables. Usually the observationsare made at discrete time steps t i l (t i ) = h( x(t i ), v(t i )) (2.3) 7
10. 10. General InformationThe vector form of this system is l (t i ) = H ⋅ x(t i ) + w(t i ) (2.4)where l (t i ) is the vector of the observations at the epoch t i , H is the observationmatrix and w(t i ) is the noise of the measurement process with the covariancematrix R (t i ) . Like the dynamic matrix, in a linear system the observationmatrix H is a constant matrix as well. 8
11. 11. Kalman Filter AlgorithmChapter 3Kalman Filter Algorithm3.1 PredictionLike mentioned before, the prediction is the first step of the Kalman filter. Thepredicted state, or better the a priori state is calculated by neglecting the dynamicnoise and solving the differential equations that describe the dynamic model − − x (t ) = F ⋅ x (t ) (3.1)The state vector at time t can be expressed by a Taylor series with respect to an −approximate state x (t 0 ) . − − − − x (t ) = x (t 0 ) + x (t 0 )(t − t 0 ) + 1 x (t 0 )(t − t 0 ) 2 + … 2by using equation (3.1) this can be rewritten as − − − − x (t ) = x (t 0 ) + F ⋅ x (t 0 )(t − t 0 ) + 1 F 2 x (t 0 )(t − t 0 ) 2 + … 2 −Thus the solution x (t ) of the differential equations, in other words the actual −predicted state is a linear combination of the initial state x (t 0 ) − − x (t ) = Φ t0 ⋅ x (t 0 ) ( 3 .2 )Φ t0 is called the state transition matrix, which transforms any initial state x(t 0 ) toits corresponding state x(t ) at time t . 9
12. 12. Kalman Filter Algorithmfrom the equations (3.1) and (3.2): − − − x (t ) = F ⋅ x (t ) = F ⋅ Φ t0 ⋅ x (t 0 ) (3.3)and by using (3.2) again, one can see − x (t ) = d − dt x (t ) = d t − dt [ ⎡d ] ⎤ − Φ 0 ⋅ x (t 0 ) = ⎢ Φ t0 ⎥ ⋅ x (t 0 ) ⎣ dt ⎦ (3 . 4 )By comparing (3.3) and (3.4) it follows that: d t Φ 0 = F ⋅Φ 0 t (3.5) dtwith the initial matrix Φ 0 = Ι , because x(t 0 ) = Ι ⋅ x(t 0 ) 0And now the covariance matrix P − (t i ) of the predicted state vector is obtainedwith the law of error propagation P − (t i ) = Φ ttii −1 ⋅ P(t i −1 ) ⋅ (Φ ttii −1 ) T + QIn the more generalized form, where also the covariance matrix of the noise Q is afunction of time, the covariance matrix is ti P − (t i ) = Φ ttii −1 ⋅ P(t i −1 ) ⋅ (Φ ttii −1 ) T + ∫ Q(t )dt (3 . 6 ) ti −1 10
13. 13. Kalman Filter Algorithm3.2 Correction −In the correction step the predicted state vector x (t i ) is improved withobservations made at the epoch t i , thus the a posteriori state has the form + − x (t i ) = x (t i ) + ∆ x(t i )with the covariance matrix P + (t i ) = P − (t i ) + ∆P(t i )As said before the Kalman filter is an optimal filter, this means that the statevariances in the state covariance matrix P + are minimized. As P − is alreadyknown from the prediction step it follows that ∆P is minimized. [ ∆P(t i ) = E ∆ x(t i )∆ x(t i ) T ]this condition is complied with ∆ x (t i ) = P − H T (HP − H T + R (t i ) ) ⋅ (l ( t −1 i − ) − H x (t i ) ) ⇒ ( ∆ x(t i ) = K (t i ) ⋅ l (t i ) − l (t i ) − )with ( K (t ) = P − H T HP − H T + R(t i ) ) −1 −K is called the gain matrix. The difference (l (t i ) − l (t i )) is called themeasurement residual. It reflects the discrepancy between the predicted − −measurement l (t i ) = H x (t i ) and the actual measurement l (t i ) .Finally the corrected state is obtained by + − ( x (t i ) = x (t i ) + K (t i ) ⋅ l (t i ) − l (t i ) − ) 11
14. 14. Kalman Filter AlgorithmIn this equation the estimated state and the measurements are weighted andcombined to calculate the corrected state. That means, if the measurementcovariance is much smaller than that of the predicted state, the measurement’sweight will be high and the predicted state’s will be low. And so the uncertaintycan be reduced.The covariance matrix of the a posteriori state is given with the law of errorpropagation by P + (t i ) = P − (t i ) − K (t i ) HP − (t i ) = (I − K (t i ) H )P − (t i ) 12
15. 15. Extended Kalman FilterChapter 4Extended Kalman Filter4.1 Nonlinear SystemsSo far only linear systems have been considered. But in practice the dynamic orthe observation model can be nonlinear.One approach to the Kalman filter for such nonlinear problems is the so-calledExtended Kalman filter, which was discovered by Stanley F. Schmidt. AfterKalman presented his results about Kalman filtering, Schmidt immediately beganapplying it to the space navigation problem for the upcoming Apollo project formanned exploration of the moon. In this process, he invented the extendedKalman filter.This Kalman filter linearizes about the current estimated state. Thus the systemmust be represented by continuously differentiable functions.One disadvantage of this version of the Kalman filter for nonlinear systems is thatit needs more time-consuming calculations. The implementation for linearsystems can be made more efficient by pre-computing the dynamic matrix F, thestate transition matrix Φ and the observation matrix H. But for nonlinear systems,these are functions of the state and consequently change with every time step andcannot be pre-computed. 13
16. 16. Extended Kalman Filter4.2 PredictionIn the nonlinear case the dynamic matrix F is a function of the state to beestimated. So the predicted state is calculated by solving the differential equationsin the form − − x (t ) = f ( x (t ))By representing this equation by a Taylor series with respect to x at the predicted −state x (t i ) and assuming that the higher order terms can be neglected, thedynamic matrix F (t i ) can be calculated with ∂f ( x) F (t i ) = ∂x x = x − ( ti )And now the other steps of the prediction can be calculated as shown in chapter3.1 with the equations (3.5) and (3.6), but it should be noted that, now the usedmatrices are not constant like in the linear case, but depend on the time step. d ti Φ ti −1 = F (t i ) ⋅ Φ ti −1 ti dt ti ∫ Q(t )dt − P (t i ) = Φ ti ti −1 ⋅ P (t i −1 ) ⋅ (Φ ti ti −1 ) + T ti −14.3 CorrectionLike the differential equations in the prediction step the corresponding nonlinearobservation equations are linearized with the Taylor series about the predicted −state x (t i ) and higher order terms are neglected.Thus the approximate observation matrix is ∂h( x) H (t i ) = ∂x x = x − ( ti ) 14
17. 17. Extended Kalman Filter −In that case the predicted measurement l (t i ) for calculating the measurement −residual (l (t i ) − l (t i )) is − − l (t i ) = h( x (t i ))Further on, we can again use the same formulas to calculate the corrected stateand its covariance matrix like in the linear case but with time dependent matrices. ( − x + (t i ) = x − (t i ) + K l (t i ) − l (t i ) )and P + (t i ) = (I − K (t i ) H (t i ) )P − (t i )with ( K (t i ) = P − H (t i ) T H (t i ) P − H (t i ) T + R(t i ) ) −1 15
18. 18. Summary Chapter 5 Summary The models and the essential implementation equations of the linear and the extended Kalman filter that were derived in section 3 and 4 are summarized below. Standard Kalman Filter Extended Kalman Filter Dynamic model: x(t ) = F ⋅ x(t ) + n(t ) x(t ) = f ( x(t ), m(t )) Observation model: l (t ) = H ⋅ x(t ) + w(t ) l (t ) = h( x(t ), v(t )) A priori state: − − − − x (t ) = F ⋅ x (t ) x (t ) = f ( x (t ))Prediction State transition matrix: d ti d ti ∂f ( x) Φ t = F ⋅ Φ ttii −1 Φ ti −1 = F (t i ) ⋅ Φ ttii −1 , F (t i ) = dt i −1 dt ∂x x = x − ( ti ) A priori covariance matrix: ti P − (t i ) = Φ ttii −1 ⋅ P + (t i −1 ) ⋅ (Φ ttii −1 ) T + ∫ Q(t )dt ti −1 16
19. 19. Summary Gain matrix: ( K (t i ) = P − (t i ) H T HP − (t i ) H T + R (t i ) ) −1 ( K (t i ) = P − (t i ) H (t i ) T H (t i ) P − (t i ) H (t i ) T + R (t i ) ) −1 ∂h( x) H (t i ) = ∂x x = x − ( ti )Correction Predicted measurement: − − − − l (t i ) = H x (t i ) l (t i ) = h( x (t i )) A posteriori state: + − ( x (t i ) = x (t i ) + K l (t i ) − l (t i ) − ) A posteriori covariance matrix: P + (t i ) = (I − K (t i ) H )P − (t i ) P + (t i ) = (I − K (t i ) H (t i ) )P − (t i ) One has to note that the parameters from the correction step are only defined at the observation times t i , whereas the parameters from the prediction step exist continuously for all t . To simplify matters, except of the state, they are calculated only at the observation times t i as well. Of course also the observations can be continuous, but in practice this is rather seldom. 17
20. 20. ImplementationChapter 6ImplementationWithin the scope of this study thesis a program in Matlab was written. Matlab-version 6.5 Release 13 was used.In this chapter the implementation is described with some illustrative sections ofthe Matlab source code.One condition was that this Kalman filter should be general. That is to say that itshould be possible for the user to determine the models. As this condition wasmore important than the computing time, the algorithm of the extended Kalmanfilter is used, because by using linear equations, it becomes just the standardKalman filter since ∂f ( x) ∂h( x) F= and H= ∂x ∂xwhereas F and H do not depend on the state in the linear case.In the following source code sections the a priori state and its covariance matrixare marked by an ‘m’ for minus and the corresponding a posteriori parameters aremarked by a ‘p’ for plus. ‘i’ specifies the time step. 18
21. 21. Implementation6.1 Extraction of the DataThe first step of the program is the reading of the required data from text files thatthe user has to prepare before using the program. How the user has to format thetext files is shown in the folder ‘format’.In this way the following data is read in:state vector * x_syminitial value of the state vector x0its covariance matrix P0differential equations of the dynamic model * fcovariance matrix of the dynamic noise * Qequations of the observation model * hobservations ltheir covariance matrices Rthe time steps of the observations tDue to their dependency on the variables of the state vector, much of this data issymbolic expressions. The concerned data is marked by *.Furthermore, the user can choose between ‘no plot’, ‘2D plot’ or ‘3D plot’ anddefine the axes.6.2 Prediction6.2.1 A Priori StateAs described in 4.2, the continuous state from the time span t i −1 to t is obtainedby solving the differential equation. In Matlab this can be realized with theordinary differential equation solvers (ode).In general, ode45, which is based on an explicit Runge-Kutta formula, is the bestfunction to apply for most problems. It is a one-step solver and for computingx(t i ) , it needs only the solution at the preceding time step x(t i −1 ) . 19
22. 22. Implementationa priori state:% solving the differential equation f to get the a prioristate% vector xm:tspan = [ti(i-1) ti(i)];[T1 X] = ode45(@diff_eq,tspan,x,opt,x_sym,f);% Each row in the solution array X corresponds to a timereturned in the column vector T1k1 = length(T1);xm = (X(k1,:)) Source code section 1with ode45 the system of differential equations dxdt = f(x(t)) is integrated overtspan from time t(i-1) to t(i) with initial conditions x, which is the +previous state vector x (t i −1 ) . Each row in the solution array X corresponds to atime returned in the column vector T1. Thus the last row of this solution array is −the a priori state xm of the time step t(i) that is x (t i ) .The system of differential equations must be coded as a function that the odesolver can use. In this case the function is called diff_eq. This function returns acolumn vector corresponding to f(x(t)).function for the differential equation:function dxdt = diff_eq(t,x,x_sym,f)% the differential equations are calculated with the valuesx:f = subs(f,x_sym,sym(x));f = sym(f);dxdt = eval(f); Source code section 2The first two arguments of the function must be t and x, even if one of them isnot used. 20
23. 23. Implementation6.2.2 State Transition MatrixTo solve the differential equation of the state transition matrix the ode45 is usedagain. But there is the problem that ode45 only works in the case of vectors andnot in the case of matrices. So we must split the state transition matrix in vectorsi.e. the differential equation is solved column by column. This is permissiblebecause matrix multiplication in general can be done column by column. Forexample each column of C, which is the solution of the matrix product C = A⋅ Bcan be calculated c j = A⋅bjwhere j is the column number.In this context, we can calculate each column of the wanted state transition matrixby solving the differential equation in the form ⎡ dΦ ⎤ ⎢ dt ⎥ = F ⋅ Φ j ⎣ ⎦jwhere j again is the number of the column.With this the ode solver is used like above. The differential equation is solved foreach column, after which the last row of the solution array Phi_trans is thecorresponding column of the wanted state transition matrix Φ (t i ) .But before solving the differential equations for the state transition matrix, thedynamic matrix F needs to be evaluated. In the nonlinear case F depends on −x(t i ) and has to be calculated at the predicted state x (t i ) like mentioned inchapter 4.2. 21
24. 24. Implementationstate transition matrix% in the case the dynamic model is nonlinear F needs to becalculated% with xmF = subs(F_dyn,x_sym,sym(xm));F = sym(F); F = eval(F)for j=1:n_s [T2 Phi_trans] = ode45(@diff_eq_P,tspan,Phi(:,j),opt,F); k2 = length(T2); Phi(:,j) = (Phi_trans(k2,:));end Source code section 36.2.3 Covariance of the A Priori StateThe covariance Q of the dynamic noise may depend on the state variables as −well, so it must be calculated at the predicted state x (t i ) , before the a prioriCovariance matrix can be computed just like shown in 4.2.Covariance matrix of the a priori stateQ = subs(Q_sym,x_sym,sym(xm));Q = sym(Q); Q = eval(Q);Pm = sym(Phi*P*Phi + int(Q,t,ti(i-1),ti(i)));Pm = eval(Pm) Source code section 4 22
25. 25. Implementation6.3 Correction6.3.1 Gain MatrixBefore the gain matrix can be computed, the observation matrix must bedetermined. In the linear case this is just the Jacobian matrix of the observationequations h , but in the nonlinear case it still depends on the state variables. Hcould simply be calculated at the a priori state like shown in 4.3. But in somecases it might enhance the results to iterate the observation matrix. Within thefollowing iteration the state, at which the observation matrix H is evaluated, isquickly improved with a simple least squares adjustment that minimizes theimprovements of the observations. So the H matrix that is calculated at thisimproved state is closer to the real observations.Iteration of the observation matrix in function ObsMatrixwhile max(abs(dx)) > 0.00001 k=k+1; % calculation of H and l0 with the (improved) initialvalues: H = subs(J,x_sym,sym(x)); H = sym(H); H = eval(H); l0 = subs(h,x_sym,sym(x)); l0 = sym(l0); l0 =eval(l0); dl = l-l0; dx = pinv(H*Rinv*H)*H*Rinv*dl; % improving the state vector x = x+dx; if k == 20 dx=[0.000001]; Iteration = 0; sprintf(Iteration in "ObsMatrix.m" aborted!nobservation matrix H will be calculated at the approximationstate xmn) endend Source code section 5 23
26. 26. ImplementationOne must note, that the standard least squares adjustment used in the iteration isnot the Kalman correction formula. As shown in chapter 3.2, with this formula thepredicted state and the measurements are weighted and combined to calculate thecorrected state. But in this iteration the intention is to calculate the observationmatrix H so that it is closer to the observations and less affected by theapproximate state. And this is realized in the iteration by minimizing theimprovements of the observations.If the observation matrix H doesn’t converge within this iteration, it is just −computed at the a priori state x (t i ) .But this iteration can pose a new problem. Namely, when not all elements of thestate vector are in the observation equations, the Jacobian matrix of h respectingthe state x has columns that are zero i.e. the matrix has no full rank and therewiththe equation system dx = ( H Rinv H ) −1 HRinv dl from the standard least squaresadjustment in the iteration has no finite solution. Hence instead of the inverse theMoore-Penrose pseudoinverse is used to solve the equation system.And then the gain matrix is computed with that observation matrix H, whichemanates from the function ObsMatrix.Gain matrixH = ObsMatrix(l(:,i),R(:,:,i),h,x_sym,xm)K = Pm*H*(H*Pm*H+R(:,:,i))^(-1); Source code section 6 24
27. 27. Implementation6.3.2 A Posteriori State and CovarianceAnd finally the predicted measurement and with it the a posteriori state and itscovariance matrix can be computed with the corresponding formulas.a posteriori state and its Covariance matrix% predicted measurement:lm = subs(h,x_sym,sym(xm));lm = sym(lm); lm = eval(lm);% measurement residual:y = l(:,i)-lm % each column of l corresponds to a time% a posteriori state vector and the corresponding covariancematrix:xp(:,i) = xm + K*y; x = xp(:,i)P = Pm - K*H*Pm Source code section 7And for the next time step xp will be the initial state x. 25
28. 28. ExampleChapter 7ExampleAs an example the program was executed to calculate the orbit of a geostationarysatellite. These geostationary orbits are used for most commercial communicationsatellites, for TV-transmissions and for meteorological satellites likeMETEOSAT. Geostationary orbits can only be achieved very close to the ring of35 786 km above the equator with an orbital speed around 11 068 km/h.7.1 Dynamic Model and Initial ValuesThe motion of an object in the earth’s gravitational field can be described by thefollowing differential equations: ⎡ x⎤ ⎡ x ⎤ ⎡nx ⎤ d2 ⎢ ⎥ GM ⎢ y ⎥ + ⎢n ⎥ y =− dt 2 ⎢ ⎥ ⎢z⎥ ⎣ ⎦ ( x2 + y2 + z2 ) 3 ⎢ ⎥ ⎢ y⎥ ⎢ z ⎥ ⎢nz ⎥ ⎣ ⎦ ⎣ ⎦where the noise elements n x , n y and n z represent for example the irregularities ofthe Earth’s gravitational field or the influence of the solar pressure.As these are second-order differential equations and the Kalman filter needs first-order differential equations, the model must be changed so that it fulfils therequirements. This can be realized by transforming one second-order differentialequation to a system of two first-order differential equations. For this purpose thestate vector x must be extended with its first derivation x . 26
29. 29. ExampleIn this example the additional components are ⎡ x ⎤ ⎡v x ⎤ ⎢ y ⎥ = ⎢v ⎥ ⎢ ⎥ ⎢ y⎥ ⎢ z ⎥ ⎢v z ⎥ ⎣ ⎦ ⎣ ⎦This causes the following equation system ⎡x⎤ ⎡v x ⎤ ⎡0⎤ ⎢y⎥ ⎢v ⎥ ⎢0⎥ ⎢ ⎥ ⎢ y⎥ ⎢ ⎥ d ⎢z⎥ ⎢v z ⎥ ⎣ ⎦ ⎢0⎥ ⎢ ⎥= +⎢ ⎥ dt ⎢v x ⎥ ⎡ x ⎤ ⎢n x ⎥ ⎢v y ⎥ − GM ⎢ y⎥ ⎢ ⎥ ⎢ ⎥ ⎢v z ⎥ ⎣ ⎦ (x 2 + y2 + z2 ) 3 ⎢ ⎥ ⎢n y ⎥ ⎢ z ⎥ ⎢nz ⎥ ⎣ ⎦ ⎣ ⎦In this way, one can reduce the form of any system of higher order differentialequations to an equivalent system of first-order differential equations.This system without the noise vector is now used to calculate the a priori statewhere G is the gravitational constant and M is the mass of the earth: m3 G = 6,6742 ⋅ 10 −11 kg ⋅ s 2 M = 5,9736 ⋅ 10 24 kgAs the computation of the actual a priori state needs the previous state vector asinitial value, one must assume values for the elements of the state vector at thestart time t 0 .For a geostationary satellite one may say that the z-coordinate is around zero andthat it stays in this range. Therewith the velocity in z-direction v z is around zeroas well. When we now assume that the satellite is situated on the x-axis at the startof its observation, also the y-coordinate is zero and the x-coordinate is thedistance from the origin, the centre of the earth. And with the mean earth’s radiusof 6370 km and the height of the satellite above equator around 36000 km thedistance between origin and satellite is around 42370 km. At this point thesatellite moves only in y-direction, therefore v y is around 11068 km/h and v x isnear zero. 27
30. 30. ExampleSo the initial state is ⎡ x0 ⎤ ⎡ 42370 km ⎤ ⎢y ⎥ ⎢ 0 ⎥ ⎢ 0⎥ ⎢ ⎥ ⎢ z0 ⎥ ⎢ 0 ⎥ x0 = ⎢ ⎥ = ⎢ ⎥ ⎢v x 0 ⎥ ⎢ 0 ⎥ ⎢v y 0 ⎥ ⎢11068 km / h ⎥ ⎢ ⎥ ⎢ ⎥ ⎢v z 0 ⎥ ⎢ ⎣ ⎦ ⎣ 0 ⎥ ⎦Let’s say that this vector is known with the covariance matrix ⎡50 km 2 ⎤ ⎢ ⎥ ⎢ 50 km 2 0 ⎥ ⎢ 50 km 2 ⎥ ⎢ ⎥ km 2 P0 = ⎢ 1 ⎥ ⎢ h2 ⎥ ⎢ km 2 ⎥ ⎢ 0 1 ⎥ ⎢ h2 ⎥ ⎢ km 2 ⎥ 1 2 ⎢ ⎣ h ⎥ ⎦and that the covariance matrix of the system noise is ⎡0 ⎤ ⎢ 0 0 ⎥ ⎢ ⎥ ⎢ 0 ⎥ ⎢ km 2 ⎥ Q= ⎢ 0,001 4 ⎥ ⎢ h ⎥ ⎢ km 2 ⎥ ⎢ 0 0,001 4 h ⎥ ⎢ 2 ⎥ ⎢ km 0,001 4 ⎥ ⎢ ⎣ h ⎥ ⎦ 28
31. 31. Example7.2 Observation Model and ObservationsFor a geostationary satellite the revolution time is T = 24 h. Only one revolutionis observed. Every three hours the distance from the origin to the satellite isobserved.Thus the observation equation is l = x2 + y2 + z2Let’s say that the distance is constantly 42156 km and its variance is 0,001 km2 sothe observation array is [42156]1x 8 km with one element of the array correspondsto one time step.7.3 ResultsIn the following table the results for one revolution are shown. For every time stepthe a priori values are faced with the a posteriori values. The units of thecoordinates are km and of the velocity km/h. ti = 3 h − x P− 30072 93.2300 60.3695 0 21.5350 25.9892 0 29926 60.3695 92.6435 0 25.9892 21.2825 0 0 0 0 32.5665 0 0 -4.5807 -7772 21.5350 25.9892 0 8.6114 6.4003 0 7860 25.9892 21.2825 0 6.4003 8.5492 0 0 0 0 -4.5807 0 0 2.1799 + x P+ 29881 16.2047 -16.2826 0 -2.2791 2.2905 0 29736 -16.2826 16.3629 0 2.2905 -2.3013 0 0 0 0 32.5665 0 0 -4.5807 -7831 -2.2791 2.2905 0 1.2487 -0.9267 0 7801 2.2905 -2.3013 0 -0.9267 1.2577 0 0 0 0 -4.5807 0 0 2.1799 29
32. 32. Example t2 = 6 h − x P− -78 23.9078 -11.0423 0 2.7856 -4.8808 0 41986 -11.0423 68.4490 0 -1.0336 26.2773 0 0 0 0 31.5209 0 0 4.4499 -11098 2.7856 -1.0336 0 1.1409 -0.9422 0 -80 -4.8808 26.2773 0 -0.9422 10.3770 0 0 0 0 4.4499 0 0 2.2151 + x P+ -106 22.1132 0.0411 0 2.6175 -0.6257 0 42156 0.0411 0.0011 0 0.0049 -0.0008 0 0 0 0 31.5209 0 0 4.4499 -11101 2.6175 0.0049 0 1.1252 -0.5435 0 -15 -0.6257 -0.0008 0 -0.5435 0.2885 0 0 0 0 4.4499 0 0 2.2151 t3 = 9 h − x P− -30046 54.5046 -9.0798 0 4.4684 -12.5291 0 29716 -9.0798 476.5877 0 -107.1142 94.2455 0 0 0 0 14.6957 0 0 -0.0584 -7823 4.4684 -107.1142 0 24.1833 -21.6615 0 -7837 -12.5291 94.2455 0 -21.6615 20.7579 0 0 0 0 -0.0584 0 0 3.4042 + x P+ -30029 47.0222 47.5445 0 -8.5441 -0.0665 0 29587 47.5445 48.0746 0 -8.6395 -0.0669 0 0 0 0 14.6957 0 0 -0.0584 -7794 -8.5441 -8.6395 0 1.5532 0.0120 0 -7866 -0.0665 -0.0669 0 0.0120 0.0005 0 0 0 0 -0.0584 0 0 3.4042 t 4 = 12 h − x P− 1.0e+003 * -42243 -350 3.4832 -4.6581 0 1.4861 0.1355 0 0 -4.6581 6.2294 0 -1.9874 -0.1812 0 67 0 0 0.0147 0 0 0.0001 -11050 1.4861 -1.9874 0 0.6341 0.0578 0 0 0.1355 -0.1812 0 0.0578 0.0053 0 0 0 0.0001 0 0 0.0034 + x P+ -42154 0.0010 -0.0023 0 0.0005 0.0001 0 -471 -0.0023 0.1132 0 -0.0114 -0.0103 0 0 0 0 14.7478 0 0 0.0865 105 0.0005 -0.0114 0 0.0016 0.0010 0 -11046 0.0001 -0.0103 0 0.0010 0.0013 0 0 0 0 0.0865 0 0 3.3928 30
33. 33. Example t 5 = 15 h − x P− -29457 1.5022 -0.9029 0 0.1126 0.3016 0 -30149 -0.9029 0.6831 0 -0.0409 -0.1852 0 0 0 0 32.8787 0 0 4.4902 7918 0.1126 -0.0409 0 0.0139 0.0219 0 -7705 0.3016 -0.1852 0 0.0219 0.0610 0 0 0 0 4.4902 0 0 2.1351 + x P+ -29469 0.6028 -0.5859 0 0.0025 0.1272 0 -30145 -0.5859 0.5713 0 -0.0021 -0.1237 0 0 0 0 32.8787 0 0 4.4902 7916 0.0025 -0.0021 0 0.0004 0.0005 0 -7708 0.1272 -0.1237 0 0.0005 0.0272 0 0 0 0 4.4902 0 0 2.1351 t 6 = 18 h − x P− 590 110.7664 29.6598 0 -8.3071 19.6877 0 -42042 29.6598 8.1447 0 -2.2026 5.3251 0 0 0 0 30.6971 0 0 -4.6430 11076 -8.3071 -2.2026 0 0.6256 -1.4708 0 203 19.6877 5.3251 0 -1.4708 3.5136 0 0 0 0 -4.6430 0 0 2.3326 + x P+ 168 3.0761 0.0470 0 -0.3140 0.3424 0 -42158 0.0470 0.0017 0 -0.0047 0.0055 0 0 0 0 30.6971 0 0 -4.6430 11107 -0.3140 -0.0047 0 0.0324 -0.0349 0 127 0.3424 0.0055 0 -0.0349 0.0385 0 0 0 0 -4.6430 0 0 2.3326 t 7 = 21 h − x P− 30077 460.8544 98.3627 0 -34.4698 -65.8387 0 -29341 98.3627 21.6793 0 -7.4826 -13.9265 0 0 0 0 48.7009 0 0 -0.5355 7778 -34.4698 -7.4826 0 2.6015 4.9014 0 7981 -65.8387 -13.9265 0 4.9014 9.4292 0 0 0 0 -0.5355 0 0 1.0338 + x P+ 30320 1.0409 1.0645 0 -0.2322 0.0059 0 -29289 1.0645 1.0906 0 -0.2378 0.0064 0 0 0 0 48.7009 0 0 -0.5355 7760 -0.2322 -0.2378 0 0.0522 -0.0014 0 7947 0.0059 0.0064 0 -0.0014 0.0004 0 0 0 0 -0.5355 0 0 1.0338 t 8 = 24 h − x P− 42392 511.0893 -159.2552 0 142.8863 -38.9394 0 767 -159.2552 50.0587 0 -44.5714 12.0875 0 0 0 0 49.1491 0 0 -0.3718 -107 142.8863 -44.5714 0 39.9526 -10.8813 0 11043 -38.9394 12.0875 0 -10.8813 2.9719 0 0 0 0 -0.3718 0 0 1.0217 31
34. 34. Example + x P+ 42148 0.0012 -0.0083 0 0.0012 0.0007 0 843 -0.0083 0.4398 0 -0.0506 -0.0456 0 0 0 0 49.1491 0 0 -0.3718 -175 0.0012 -0.0506 0 0.0062 0.0052 0 11062 0.0007 -0.0456 0 0.0052 0.0050 0 0 0 0 -0.3718 0 0 1.0217The biggest difference between the predicted and the corrected state occurs at thetime step t 6 = 18 h. At this time step it is around ⎡ − 422 km ⎤ ⎢ − 116 km ⎥ ⎢ ⎥ ⎢ 0 km ⎥ ∆x(t 6 ) = ⎢ ⎥ ⎢ 31 km / h ⎥ ⎢− 76 km / h⎥ ⎢ ⎥ ⎢ 0 km / h ⎥ ⎣ ⎦At the time step t 5 = 15 h the improvements are the smallest with ⎡ − 12 km ⎤ ⎢ 4 km ⎥ ⎢ ⎥ ⎢ 0 km ⎥ ∆x(t 5 ) = ⎢ ⎥ ⎢ − 1 km / h ⎥ ⎢− 2 km / h⎥ ⎢ ⎥ ⎢ 0 km / h ⎥ ⎣ ⎦Furthermore it is shown how the changes, made by the observations in thecorrection step, affect the accuracy. If the covariance matrix of the observations issmaller than that of the predicted state, the states covariance matrix is improved asone can see in the table. 32
35. 35. ExampleThe following figure shows the calculated trajectory of the geostationary satellite.The red crosses mark the corrected states at the time steps, where the observationsare made. Figure 2: calculated orbitTo pinpoint the correction in a graphic display the plot around the state at the timestep t 6 is enlarged. Figure 3:Enlargement 33
36. 36. Conclusion and ProspectChapter 8Conclusion and ProspectThe program that was written in the scope of this study thesis intends to provide abasic understanding of the Kalman filter. Of course this program is too slow forreal time computations, because the handling with symbolic expressions in Matlabneeds a lot of calculation time.But in general the recursive nature of the Kalman filter allows for efficient real-time processing. Because of this and because of its apparently simple and easilyprogrammed algorithm, the Kalman filter will continue to play a very importantrole not only in GPS-based navigation systems. 34
37. 37. ReferencesB. Hoffmann-Wellenhof, H. Lichtenegger and J. Collins (1997): GPS Theory andPractice, Springer-Verlag ViennaMartin Vermeer (2004): Methods of Navigation, Lecture notes, department ofsurveying of Helsinki University of Technology(http://www.hut.fi/~mvermeer/nav_en.pdf)Mohinder S. Grewal, Angus P. Andrews (2001): Kalman Filtering: Theory andPractice Using Matlab, John Wiley & Sons Inc 2001Sorenson, Harold W. (1985): Kalman filtering: theory and application, New YorkInternet:http://www.wikipedia.dehttp://www.cs.unc.edu/~welch/kalman/ 35